The AI Research Agent Prompt Most People Get Wrong

Most AI research agent guides are wrong about what makes a research agent good. They obsess over output polish - tone, structure, executive-summary formatting - when the thing that actually matters is whether a single claim in the report can be trusted. A beautiful report full of invented citations is worse than a rough one that's true.

An AI research agent should be judged like a journalist, not a copywriter. This teardown takes a common prompt that produces gorgeous fiction and rebuilds it into one that grounds every claim in something real.

Before: The Weak Prompt

Here's the prompt that shows up in most tutorials:

You are an expert research assistant. Research the topic below and write a
comprehensive, well-structured report with sections, insights, and a
conclusion. Be thorough and authoritative.

Topic: {topic}

What this does: it asks the model to write an authoritative report from its own memory, with zero requirement that any statement trace back to a source.

Paired with a web-search tool, this looks like it works. The report has headings, confident numbers, tidy conclusions. Ship it a few times and you'll notice the pattern: some of those confident numbers are wrong, and a couple of the "sources" it references don't say what the report claims - or don't exist.

Why the AI Research Agent Prompt Fails

The prompt optimizes for the wrong target. "Comprehensive and authoritative" tells the model to sound like a finished report, and a language model is extremely good at sounding finished. Authority is a style, and the model will happily produce the style whether or not the substance is there.

Nothing in the prompt separates "things I found" from "things I'm asserting." So the model blends retrieved facts and plausible-sounding filler into one smooth voice, and you - the reader - can't tell which sentence came from a real source and which came from the model's priors. The failure isn't that the model lies on purpose. It's that the prompt never gave it a job where honesty and helpfulness point in the same direction.

⚠️ Common mistake: assuming a web-search tool makes an agent factual. Search gives the agent access to facts; it doesn't force the agent to use them. Without a prompt that binds claims to retrieved text, the agent will search, skim, and then write mostly from memory anyway - decorating it with a citation or two for cover.

After: The Improved Prompt

The rewrite changes the job from "write a report" to "assemble a report where every non-obvious claim carries a source, and gaps are labeled as gaps."

You are a research agent. Work in two phases.

PHASE 1 - GATHER: Use the search tool. For each useful finding, record it as
  {"claim": ..., "source_url": ..., "quote": "<=25 words supporting it"}.
  Do not write prose yet.

PHASE 2 - WRITE: Draft the report using ONLY claims from Phase 1. After each
  factual sentence, cite the source_url. If a needed fact was never gathered,
  write "[GAP: not found in sources]" instead of guessing. Do not add claims
  that lack a Phase 1 record.

Topic: {topic}

What this does: it splits gathering from writing, requires a short supporting quote per finding, and forces the agent to mark missing facts as explicit gaps rather than invent them.

The two-phase split is the entire trick. By making the agent commit its evidence before it's allowed to write, you break the habit of writing first and back-filling citations. And the

[GAP]

Breaking Down Each Element

The

quote

The evidence-first ordering matters because generation is autoregressive. Once the model starts writing confident prose, it's committed to continuing confidently. Gathering evidence into a structured list first gives the writing phase a fixed set of raw materials it can't quietly expand.

There's a subtler benefit to the structured Phase 1 record. Because each finding is a discrete object with a claim, a URL, and a quote, you can programmatically check the report against it. A short script can confirm that every cited URL in the final draft appears in the Phase 1 list, and flag any sentence that asserts a fact without a citation. That turns "trust the agent" into "verify the agent," which is the only footing on which an AI research agent belongs anywhere near real decisions. You wouldn't accept an unsourced number from a junior analyst; the same standard should apply to a model, and the two-phase structure is what makes enforcing it cheap.

One caveat worth stating plainly: this pattern reduces fabrication, it doesn't eliminate it. A determined model can still misread a source or pull a quote out of context. The verification pass and the human spot-check aren't optional extras - they're the second and third layers of a defense that no single prompt can provide alone.

The

[GAP]

⚡ Pro tip: render each cited

source_url

Variations for Different Contexts

A market analyst researching a competitor adds a recency constraint to Phase 1 - reject sources older than 12 months - because stale funding numbers are worse than no numbers. The gap markers then flag where current data simply doesn't exist publicly, which is itself a finding.

A policy researcher requires two independent sources per contested claim and has the agent note when its two sources disagree, surfacing controversy instead of flattening it into false consensus. This turns the AI research agent from a summarizer into something closer to a fact-checker.

A product manager doing competitive research points the agent at specific docs and changelogs rather than open search, so every claim traces to a primary source the team can verify in one click. Primary-source-only is the strictest and safest mode - use it whenever a wrong claim would embarrass you in front of leadership.

The tradeoff across all these modes is speed against trust, and it's a real one. Two-phase gathering, quote requirements, and a verification pass together make the agent slower and pricier per report than the naive "write it all at once" prompt. That cost is the point. You're buying the ability to hand the output to someone who will act on it without re-doing the research themselves. For a throwaway internal skim, the loose prompt might be fine. For anything that informs a decision, a hire, an investment, or a public claim, the extra minutes and tokens are the cheapest insurance you'll ever buy against confidently publishing something false.

A technical writer documenting an API takes the strictest stance of all: the agent may only cite the official reference and the changelog, and any behavior not documented there gets a

[GAP]

The two-phase skeleton stays; the constraints on Phase 1 change with the stakes. What never changes is the core discipline: gather first, cite everything, and make ignorance visible. A report you can audit beats an eloquent one you have to take on faith, every single time.

⚡ Pro tip: for anything high-stakes, run the finished report through a second agent whose only job is to verify each citation against its quote and flag mismatches. A cheap verification pass catches the failures a busy human reviewer skims past.

Save and Reuse This

The gathered-evidence-then-write pattern is the reusable asset here, not any single report. Once you've tuned the phases and constraints for your domain, that prompt is worth protecting.

⚡ Pro tip: keep a per-domain version of the Phase 1 constraints - recency for markets, source count for policy, primary-only for product. Small tweaks, big reliability differences, and you'll forget which is which within a month if you don't write them down.

Store your research-agent prompt and its domain variants in a shared library like PromptABCD so the source-grounded version is the one your whole team reaches for by default. A research agent is only as trustworthy as its prompt, and the last thing you want is a teammate quietly running the "be thorough and authoritative" version on a report headed for the board.

Treat the prompt itself as a living document. Every time the verification pass catches a new failure mode - a source type the agent over-trusts, a claim category it tends to invent - fold the fix back into the shared version so the whole team inherits the lesson. A research agent that gets more careful every month is one you can gradually trust with higher-stakes work; one frozen at version one stays exactly as fallible as the day you wrote it. The prompt is the product here, and like any product it earns its reliability through iteration rather than a single clever draft.